Abstract: An approach called constrained random selection for relevance feedback is proposed in this paper.At first,all the images are sorted by similar measure,and then a threshold is selected to restrict the space of random selection.At last,the restricted space is divided into some sub-spaces,and random selection is applied to these sub-spaces to enlarge the training sets and resolve the small sample problem preferably.In addition,we compute the weights of multiple SVM classifiers dynamically and fuse the single results to resolve the users' preference problems in relevance feedback preferably.Experimental results demonstrate the effectiveness of the method.